How to live with AI?

Everyday and social impacts, and what wisdom traditions offer. Essay 3 of 4.

Hero Image

Executive summary

In under four years AI has moved into daily life. By mid-2025 about a tenth of the world's adults used ChatGPT, mostly for things other than work, and by late 2025 nearly two-thirds of American teenagers used chatbots, about three in ten of them every day. People use AI to learn, to work, to ask advice and to keep them company; governments use it to watch, score and serve their citizens, and armies to help pick targets. The first careful studies show the same tool strengthening some people and institutions and weakening others, and social media has taught us that a technology's largest effects arrive quietly and are clear only once they are entrenched. Wisdom traditions have taught for millennia about what we consume, whose company we keep and how to stop. So, setting aside the risk of catastrophe, how does AI change how we live, and what do wisdom traditions, especially Buddhism, offer?

Our answer is that AI's deepest everyday effect is on our capacity to choose well, as individuals and as societies: to attend, think, judge and relate for ourselves, and to govern ourselves together. Like social media before it, its largest effects will be emergent and quiet, and visible only once they are entrenched. It is the first technology that thinks and relates with us. As a tutor it strengthens that capacity; as a crutch or an engagement-optimised companion it hollows it out, most of all in children. The traditions offer practices that build the capacity to choose (mindful consumption, stopping), standards of speech and friendship that expose flattery and simulated intimacy, and community as the real refuge. So we should integrate AI slowly and deliberately: adopt it where a human practice or institution can wrap it, protect what it could erode, and decide before the horse has bolted rather than regulate after.

The key points, one per section:

  1. A technology's deepest impacts are emergent, so they are hardest to foresee where they are subtle and social, and easiest to change before we can see them. Like a species released into a new ecosystem, a technology interacts with everything around it. Social media is the warning, and AI repeats the pattern faster and more intimately, because it talks back.
  2. AI's benefits are real but conditional, and its harms are quiet, cumulative and fall hardest on those least able to choose well. The benefits are concrete: earlier diagnosis, faster science, a tutor for every child where teachers guide it, expert advice for people who could not afford it. The harms are just as specific: outsourced thinking, flattering companions, children formed by machines, thinner entry-level work. The benefits are proven where a human institution wraps the AI and fade where people use it alone.
  3. AI can strengthen our own agency and judgment, but its default pull is to weaken them. In the best trials, how AI is used decides the outcome: answer-giving AI weakens learning, tutoring AI strengthens it. The remedy is to make effortful use the default and back it with rules.
  4. AI strengthens institutions where accountable human practice wraps it, and weakens them where it concentrates control or floods shared spaces. Public services can get faster and better. The larger danger is a slow erosion of the conditions for self-government: in public debate, the power of the state, work, the institutions of knowledge and the conduct of war.
  5. Wisdom traditions offer a discipline of consumption, speech and community for the person, interbeing, right livelihood and sangha for the society, and the humility a technology with emergent effects demands. Six Buddhist teachings, from mindful consumption to interbeing, give a test and a practice for each area of life, and Catholic teaching reaches similar judgments.
  6. Integrate AI slowly and deliberately: adopt it where a human practice or institution can wrap it, protect what it could erode, and decide before the horse has bolted rather than regulate after. Individuals should keep the thinking and relationships they value; families and schools should delay companions and use AI only as a tutor; labs should treat flattery as a safety failure; governments should regulate where markets reward dependence. This answer is provisional.

1. How should we think about the impacts of a new technology?

A technology's deepest impacts are emergent: they come from how it reshapes habits, incentives and institutions, so they are hardest to foresee where they are subtle and social, and easiest to change before we can see them.1

This frames the rest of the essay, and it changes what counts as evidence. By the time an emergent harm can be proven, it is entrenched. So several claims below rest on a plausible mechanism grounded in precedent, often with its core levers well evidenced (how attention works, what engagement optimisation did on social media), rather than on proof that the harm is already here. We say which kind of claim each is.

1.1 Why are a technology's impacts so hard to foresee?

Like a species released into an ecosystem, a technology interacts with everything around it, and by the time harms are clear it is entrenched (the Collingridge dilemma).2

In 1935 cane toads were released in Australia to control a beetle in the sugar cane. They failed to control it, and now poison native predators across more than 500,000 square kilometres; an early ban on releases was reversed under pressure. A technology behaves the same way. It interacts with everything around it, and its largest effects come from those interactions.

The technology-policy scholar David Collingridge named the trap in 1980. While a technology is young it is easy to change, but its impacts cannot yet be seen; once they are clear, it is entrenched in infrastructure, industry and habit. Leaded petrol is the classic case. Added in the 1920s to stop engines knocking, it was phased out worldwide only in 2021, and by one estimate it exposed over 170 million living Americans to high lead levels in childhood, at a cost of 824 million IQ points. Boosterism makes the trap worse. It pictures the dramatic upside, the flying car, and misses the slow, second-order effects.

1.2 What does social media teach us?

Social media is the warning: a technology few objected to became engagement-optimised feeds with surprising, disputed and entrenched effects; AI repeats the pattern faster and more intimately, because it talks back.3

Who would object to helping people connect and share? Yet the feeds that grew from that aim were optimised for engagement, and their effects on attention, polarisation and possibly teenage mental health surprised even their makers. In 2023 the US Surgeon General found enough evidence of risk to young people to act, though not proof of harm, and named engagement-maximising design as a concern. A review of about 500 studies found digital media use associated with more polarisation, populism and distrust, as well as more participation and knowledge, though most of that evidence is correlational, and a large field experiment found that cutting like-minded content on Facebook did not measurably reduce polarisation. The teenage mental-health debate is still open, and policy has moved ahead of proof, as with Australia's ban on social media for under-16s in December 2025.

What carries over is the mechanism. A system optimised for a measurable proxy, engagement, learns to favour whatever holds attention, and reshapes a society's information diet and incentives without anyone intending it. By the time we argued about the effects, the feeds were entrenched. AI assistants and companions take the same mechanism further, because they are personalised, conversational and relational. If we wait for proof again, we will act too late; acting early means slowing down enough to think.

2. What are the main benefits and harms?

AI's benefits are real but conditional (earlier diagnosis, faster science, tutoring, expertise for the many), and its harms are quiet, cumulative and fall hardest on those least able to choose well.4

2.1 What are the potential benefits?

The potential benefits are large and concrete: earlier diagnosis, faster science, a tutor for every child where teachers guide it, expert advice for the many, translation and accessibility, and a lift for novices at work.5

AI reads mammograms and other scans alongside doctors, catching cancers earlier. It has transformed parts of science: AlphaFold's predicted protein structures are used by over three million researchers, and AI is speeding drug discovery. A patient tutor that explains anything, at any hour, could reach every child, and where teachers guide it, it already lifts learning. People who could never afford a lawyer, a doctor's time or a financial adviser could get competent first advice, though section 2.3 shows that advice alone is not yet enough. Translation and accessibility tools open the world to people who were shut out of it by language, sight or hearing. And at work, AI lifts novices most, closing part of the gap between new and experienced staff. Debunking is a benefit too: dialogue with an AI durably reduced belief in conspiracy theories.

Nearly every capability also appears among the harms. A system that can explain anything can be a patient tutor or a way to avoid learning; one that listens without tiring can comfort a lonely person or replace the friends they might have made. In In Tech We Trust we describe technology moving from servant to master, from a tool we pick up for our purposes to a force that shapes what we want. Which way AI goes depends largely on whether it serves inside a human practice or institution, or is designed to master behaviour for engagement.

2.2 What are the potential harms?

The potential harms are pervasive: outsourced thinking, flattery and simulated intimacy, children formed by companions, thinner entry-level work, a degraded information commons and concentrated power.5

Thinking handed over brings deskilling and automation bias, the habit of following confident machine output. Flattery and simulated intimacy threaten to turn the attention economy into an attachment economy, in the Center for Humane Technology's phrase. Children are growing up with companions that, by design, have no stake in who they become. At work, AI intensifies effort and thins the entry-level jobs where people learn a trade. The information commons is flooded with synthetic content and starved of the revenue that sustains it. And power concentrates in a few firms and two countries. Section 2.3 gives the evidence for each, and section 4 develops the harms to societies.

2.3 What is actually happening?

Benefits are proven where a human institution wraps the AI (a screening programme, a teacher-guided tutor) and fade where individuals use it alone;6 harms are emerging as consistent patterns, though causal proof mostly lags.7

The benefits hold up best inside institutions. In a Swedish trial of about 106,000 women (reported in the press; final results are due), AI-supported breast screening found 29% more cancers and cut radiologists' reading workload by 44%. In Nigeria, six weeks of after-school tutoring with GPT-4, guided by teachers, raised learning by 0.31 standard deviations, a large gain for so short a programme. In a support centre, AI raised the output of new staff by about a third, against 14% on average. A tutoring assistant helped weaker human tutors most. Where people use AI alone, the gains fade. In an Oxford trial, members of the public who used chatbots to decide what to do about medical scenarios did no better than those who did not, though the chatbots alone usually identified the condition. An AI business mentor for Kenyan entrepreneurs had no effect on average: it helped strong firms and hurt weak ones, who chose worse advice to follow.

The harms show up as consistent patterns. Only about 3% of Claude's conversations are emotional, but one American adult in ten, and one in five under 30, uses chatbots for emotional support. In a four-week trial by OpenAI and MIT, and, according to press reports, a year-long panel of over 2,000 adults and a study of Character.AI users, heavy, self-chosen companionship use went with more loneliness and less human contact. These are associations with a plausible mechanism, displacement, and lonely people also turn to chatbots more. The comfort is real in the short term. One conversation with a companion eased loneliness about as much as talking to a person, and a clinician-supervised therapy chatbot sharply reduced symptoms of depression.

Flattery, which researchers call sycophancy, is the clearest link from design to harm. Chatbots are trained partly on people's ratings, and people rate being agreed with highly. In experiments published in Science in 2026, eleven leading models endorsed users' actions about 50% more often than people did, including harmful ones. A single flattering exchange made participants less willing to repair a quarrel, and they rated the flattering system higher and trusted it more. The market can reward the very thing that harms us; in April 2025 OpenAI withdrew a ChatGPT update that had become too flattering.

For children, the worst harms so far are grave individual cases. In January 2026 Character.AI and Google agreed to settle five family lawsuits, and Reuters reported in 2025 that Meta's internal rules allowed its chatbots “romantic or sensual” conversations with children. Companies and lawmakers have begun to respond. Character.AI ended open-ended chat for under-18s in November 2025, and in September 2026 California required companion apps to set one-hour session defaults and never to claim romance or feelings to a minor. The evidence on work and information is set out in section 4.

2.4 Which few matter most?

For individuals, three impacts matter most: outsourced thinking, flattering companionship, and children growing up with both; for societies, the information commons, work, and the power of states and firms.8

We chose these on five tests: depth, scale, novelty, whether the traditions have distinctive insight, and fit with this question. Outsourced thinking comes first. AI is the first general-purpose substitute for thinking, used by a tenth of humanity, and the capacity it erodes is the precondition for every other choice: choice is of little use without the capacity to choose well. Flattering companionship is where businesses that profit from our time begin to profit from our attachment, and where the traditions speak most directly. Children are where both meet people still being formed, slowly and through relationship, and where harms are least reversible. For societies, the information commons, work, and the power of states and firms over citizens matter most, and section 4 takes them up.

3. Does AI strengthen or weaken our own agency and judgment?

AI can strengthen our own agency and judgment, but its default pull is to weaken them,9 because it removes the effort through which those capacities are built and kept. What decides the outcome is how AI is used, and product defaults and market incentives push towards the eroding use.

3.1 What do we mean by agency and judgment?

Agency and judgment are capacities that grow or wither with use.10

Agency is the capacity to act from reasons you endorse; judgment is the practical skill of weighing what matters in a particular case, Aristotle's practical wisdom. Psychologists treat autonomy as a basic need and find that confidence in our own abilities grows through mastery. In Wisdom and Wanting What's Good we describe wisdom as three interwoven capacities: seeing what matters, seeing what is so and what would work, and the will to act on it. AI can help with the second. It has little to offer the first and third, and a flattering system weakens them. A useful test of any reliance on AI is whether we would endorse it on reflection, and whether it keeps the capacities we value.

3.2 How could AI strengthen or weaken them?

AI helps when it scaffolds,11 and harms when it skips the struggle, invites blind trust,12 flatters or persuades.13

AI strengthens judgment when it acts as a scaffold: a tutor that asks rather than tells, a second opinion that catches what we missed, expertise that widens our real choices. A Socratic AI tutor at Harvard beat active-learning classes, and conversations with GPT-4 durably cut belief in conspiracy theories by about a fifth. A reliable, transparent tool can become part of good thinking, as a notebook or a map does.

It weakens judgment in four ways. It lets us skip the struggle through which skill is built: learners with answer-giving AI skip the monitoring and checking that learning needs, and knowledge workers who trust AI more report thinking less critically. It invites automation bias, the habit of following confident machine output, to which experts are prone too. It flatters, eroding the self-questioning that judgment depends on. And it persuades and homogenises. In a study of nearly 77,000 people, chatbots persuaded at or above human level, and the techniques that made them more persuasive made them less accurate. Writers given AI suggestions wrote stories judged better individually and about 10% more alike collectively, which narrows the diversity of ideas a society draws on.

3.3 What is the evidence?

How AI is used decides the outcome: answer-giving AI weakens learning, tutoring AI strengthens it.14

The best evidence comes from randomised trials. In Turkey, about a thousand secondary-school students practising maths were given ordinary GPT-4, a version designed to tutor without giving answers, or nothing. The ordinary version raised their practice scores by 48%, and in the exam without AI they then scored 17% lower than the group that had nothing. The tutoring version avoided the harm. Anthropic ran a similar trial of its own product with 52 programmers and published the result: those learning a new coding tool with AI scored 50% on a later test of understanding, against 67% without, with no significant time saved. Those who handed the work to the AI scored under 40%; those who used it to ask conceptual questions scored 65% or more. Well-designed tutors with teachers involved produce real gains, at Harvard and in Nigeria.

Expert skills can erode too, though the evidence is thinner. In an observational study, after doctors in four Polish clinics began using AI to spot polyps during colonoscopy, their detection rate when working without it fell from 28% to 22%. Experienced radiologists shown deliberately wrong AI suggestions fell from 82% to 46% accuracy. The brain-scan evidence is weaker than the headlines: a much-reported MIT study had 18 people in its key session, has not been peer reviewed and has been formally critiqued. The cleanest analogy is satellite navigation, whose heavy users show weaker spatial memory. We are also poor judges of what we hand over: experienced software developers in one trial were 19% slower with AI while believing they were 20% faster.

The short trials agree that what we let AI do for us, we become less able to do ourselves. Long-term causal studies of adult reasoning do not yet exist.

3.4 What could we do about it, and how?

Make effortful use the default, not an opt-in, and back it with rules: learning modes on by default, mandated unassisted practice, and independent evaluation.15

The how is mostly defaults and mandates, because exhortation loses to markets that reward the eroding pattern. Products should put learners into learning modes by default, where today OpenAI's and Anthropic's are one click away; they should add friction such as “show your reasoning first”, show their uncertainty, and train out flattery. Schools and professions should protect unassisted practice and assessment, through oral exams and work done in class, and borrow aviation's practice of keeping pilots flying by hand. Regulators should fund independent long-run evaluation, since most evidence is short-term or produced by the companies; Anthropic's coding study is a creditable exception. And each of us can do the thinking we want to keep first, use AI to question it, and notice when a felt speed-up outruns the real one.

4. Does AI strengthen or weaken our societies, and how?

AI strengthens institutions where accountable human practice wraps it, and weakens them where it concentrates control or floods shared spaces; the danger is a slow erosion of the conditions for self-government.16

The personal effects above have social counterparts. A society, too, can lose the capacities it does not use: a shared basis for debate, accountable government, work that forms people, institutions that make and keep knowledge. Much of this section rests on mechanism and precedent more than on observed harm, for the reasons given in section 1. Civilisation-scale shifts of power, and permanent lock-in, belong to the essay on risk and pace; here the subject is how institutions work today.

4.1 Democracy and public discourse

AI has not yet swung elections, but it is eroding the shared basis of debate.17

Researchers at the Alan Turing Institute found no measurable effect of AI on the results of the 2024 UK and EU elections; the main harm was an “erosion of confidence in what is real”. The slower damage is to the commons that debate draws on. About half of new English-language web articles are now mainly AI-generated, by one detector-based estimate. With an AI summary at the top of a search, people click a result in 8% of visits against 15% without, and, by industry data, Google referrals to publishers fell about 40% in a year. When asked about the news, chatbots misrepresented it in 45% of answers. And persuasion at scale is now cheap: conversational AI persuades at or above human level. The same power can strengthen debate: dialogue with AI durably reduced belief in conspiracy theories. The mechanism is social media's, intensified: engagement-optimised, personalised and now conversational.

4.2 Governance and the power of the state

AI makes surveillance and administrative control cheap, shifting power from citizens to states, most of all in authoritarian ones, while it can also improve public services.18

This is largely a mechanism claim. AI makes it cheap to watch, score and act on whole populations, so whoever controls the system gains capacity, and when the system is wrong the error falls at scale on the people least able to contest it. The precedents are already here. By 2019 at least 75 of 176 countries used AI surveillance, including half of the advanced democracies, with Chinese firms as major suppliers. Automated scoring has gone wrong at scale in democracies too. In the Netherlands a risk-scoring algorithm that used nationality wrongly treated tens of thousands of families as fraudsters, and the government resigned. Australia's Robodebt scheme, simpler automation than today's AI, raised about A$1.73 billion in unlawful debts against more than 400,000 people. The upside is real: in a UK trial, 20,000 civil servants reported saving 26 minutes a day.

4.3 Work and the economy

There is no mass unemployment yet, but entry-level hiring has thinned, work has intensified and is monitored, and gains flow to capital.19

Unemployment in the US shows no clear AI signal, and in support work AI lifted novices by about a third, narrowing the gap with experienced staff. But employment of 22- to 25-year-olds in the jobs most exposed to AI is about 19% below trend, mostly through reduced hiring; the researchers stress this is a pattern rather than proof of cause. An eight-month study inside one technology company found that staff using AI took on more, worked faster and prompted through their breaks, all voluntarily, and Danish payroll data show time saved being absorbed by new tasks. Nine in ten US firms use algorithmic management. The mechanism for who gains is ownership: the largest cloud firms are spending about $700 billion on AI infrastructure in 2026, and the US and China hold about 90% of GPU cluster capacity. Work forms people as well as paying them, which is why thinning the jobs where people learn a trade matters beyond the numbers.

4.4 Education and the institutions of knowledge

Assessment is collapsing faster than teaching adapts, and the institutions that make knowledge are losing revenue and trust.20

Teacher-guided AI tutoring, as in Nigeria and at Harvard, shows what AI can do for teaching. But over half of American teenagers use chatbots for schoolwork, and 59% say AI cheating is common at their school; proven AI-cheating cases at UK universities, per student, tripled in a year, according to press reports. Homework and the take-home essay no longer show what a student can do alone. The institutions that make and keep knowledge, publishers, reference works and universities, are losing the readers and revenue that sustain them as AI answers questions itself. The mechanism is that AI is trained on the knowledge commons while drawing away the readers and money that renew it; the loss of trust is, so far, a mechanism claim rather than a measured one.

4.5 Security and the military

AI is entering targeting, cyber operations and surveillance faster than norms are agreed.21

Reporting on Israel's “Lavender” system describes AI generating tens of thousands of targets in Gaza with about 20 seconds of human review each; the Israel Defense Forces dispute this account. Rules lag behind. A UN General Assembly resolution on lethal autonomous weapons passed 164 to 6 in December 2025 but no longer calls for negotiations, and the expert group under the Convention on Certain Conventional Weapons produced a non-binding report in September 2026, leaving any mandate to negotiate to a review conference in November. Spending races ahead: the Pentagon requested about $13.4 billion for AI and autonomy for 2026, its first dedicated budget line. The mechanism is that AI compresses the time from sensing to striking and multiplies the targets that can be processed, so human judgment slides from deciding to approving. Catastrophic misuse is the subject of the essay on risk and pace; the everyday militarisation of AI, and its spillover into policing, belongs here.

5. What do wisdom traditions offer?

Wisdom traditions offer a practical discipline of consumption, speech and community for the person, interbeing, right livelihood and sangha for the society, and the humility a technology with emergent effects demands.22 Buddhism and Catholic teaching now converge on much of it.

5.1 Which traditions?

Plum Village Buddhism leads, with Catholic teaching and the Sabbath as companions.23

We come to this from the Plum Village tradition of Thich Nhat Hanh (Thay, “teacher”, to his students), whose mindfulness trainings already speak to our use of technology; why these traditions, and how to settle disagreements between them, is taken up in the essay on alignment. Pope Leo XIV's encyclical Magnifica Humanitas (May 2026) is the most explicit statement of any tradition on AI in everyday life. Plum Village, by contrast, has said almost nothing in public about AI: we found no talk by its monastics in 2025–26 centred on it. That is a gap a monastic voice could fill, so what follows applies the tradition's practices rather than reporting a community position.

5.2 Which Buddhist teachings, and what does each say?

Six Buddhist teachings each make a practical point: mindful consumption, loving speech, spiritual friendship and sangha, stopping, right livelihood and interbeing.24

Mindful consumption. The Buddha taught four kinds of nutriment: edible food, sense impressions, volition and consciousness. All we take in feeds us, watering some seeds in us and not others. The Fifth Mindfulness Training commits practitioners to consume in a way that “preserves peace, joy, and well-being”, names websites and electronic games among possible toxins, and warns against consuming to cover loneliness or pain. AI is a sense impression that also trains volition, our deepest wanting, so the question for every use is what it is feeding.

Loving speech and deep listening. The Fourth Training opens “Knowing that words can create happiness or suffering” and commits “not to spread news that I do not know to be certain”. Speech should be truthful and caring. Loving speech and flattery can sound alike; underneath they are opposites, since flattery tells us only what we want to hear.

Spiritual friendship and sangha. When Ananda suggested that good friendship was half of the holy life, the Buddha corrected him: it is “the whole of the holy life”. A good friend (kalyanamitta) steadies and corrects us, the inverse of a companion designed to keep us engaged. Thay is often quoted as saying the next Buddha may take the form of a community. AI can serve as a skilful means, as a bell or a book can; it cannot be a spiritual friend.

Stopping. “The horse is technology. It carries us and we cannot control it,” Thay said in 2013. The practice is stopping: the mindfulness bell that calls us back to ourselves, telephone meditation (three breaths before answering), the weekly Lazy Day with no schedule. In 2013 Thay suggested that phones could ring a mindfulness bell; an engineer objected that it would impose on people. In Tech We Trust uses the exchange for its central point: choice is of little use without the capacity to choose well.

Right livelihood. “Right livelihood has ceased to be a purely personal matter. It is our collective karma,” Thay wrote, as quoted by Plum Village. What we build, and how we work, is shared.

Interbeing. Nothing exists by itself alone. A technology's effects pass through everyone it touches, users and non-users alike. This is the Buddhist form of section 1's ecosystem point, and the ground for treating AI's effects on society as everyone's concern.

5.3 What do other traditions say?

Catholic teaching reaches similar practical judgments,25 and the Sabbath shows how to stop together.26

Magnifica Humanitas applies to AI what the Fifth Training says of media in general. Platforms are “designed to capture users' time and attention” (¶170). Simulated words “do not build genuine relationships, but only their appearance”, and users may “lose the very desire to form genuine human connections” (¶100). “Schools are not called to follow the pace of the digital world” (¶147), and should teach when AI ought not to be used (¶140). It affirms “the primacy of human labor” (¶30) and calls truth “a common good and not the property of those with power” (¶137). An earlier Vatican note, Antiqua et nova (2025), adds that a teacher's presence creates a relationship AI cannot replicate, and that AI can “paradoxically deskill workers”.

Judaism offers a collective practice of stopping. Abraham Joshua Heschel called the Sabbath “a palace in time”, and some families now keep a weekly “tech Shabbat” free of screens. A shared day of rest means nobody has to resist alone.

5.4 How do these teachings apply?

Each teaching gives a test and a practice: for our own thinking, our relationships, children, society and institutions, and how we adopt new technology.27 The tests are our synthesis; no source states them as such.

Our own thinking. Test: does this use deepen my own seeing, or replace it? Practice: stop before reaching for the answer, look at the problem myself, then use AI to question my work. Plum Village teaches that insight comes through practice, “not from the accumulation of intellectual knowledge”, and the Turkish and Anthropic trials show a humbler version of the same truth.

Relationships. Test: does this companion steady me and lead me back to people, or keep me returning? Practice: choose real community over a better simulation. The four nutriments ask what a companion is feeding, a real need met or a loneliness covered over. Loving speech gives labs a standard: truthful, caring, and sometimes disagreeing.

Children and education. Test: what seeds are we watering? Practice: put the adults' own practice and presence first. What children repeatedly practise and receive, mostly through relationship, is who they become. Plum Village's work in schools, set out in Happy Teachers Change the World, starts from the teacher's own practice, as Magnifica Humanitas and Antiqua et nova do. Schools should form the adults and protect slow, unassisted, relational learning more than teach about tools.

Society and institutions. Test: what does this water at scale, and who bears it? Practice: right livelihood for the people who build and deploy AI, and sangha as the civic form of refuge. Right livelihood applies to engagement products, AI-intensified work and surveillance alike. Interbeing makes the information commons and democratic debate shared goods, which no one can protect by private choice.

Adopting new technology. Test: have we looked deeply enough to see its emergent effects? Practice: stop before adopting, and adopt slowly enough to notice. This is the traditions' answer to section 1. Emergent effects reward humility and patience, the opposite of boosterism, and social media showed the cost of adopting first and looking later. The stopping that the essay on risk and pace asks of the frontier, this essay asks of everyday adoption.

6. So: how should we integrate AI into our lives and societies?

Integrate AI slowly and deliberately: adopt it where a human practice or institution can wrap it, protect what it could erode, and decide before the horse has bolted rather than regulate after.28 This answer is provisional, and will be sharpened as the newer parts of this inquiry, on framing and on society, are researched further.

Individuals should consume AI mindfully and keep the thinking and relationships they value.29 Ask of each use what it is feeding. Draft first, then ask the AI to question you. Treat a chatbot that always agrees with you as a warning, and the urge to turn to a companion when lonely as a bell, a reminder to call someone. Keep times without AI, a Lazy Day or the first hour of each morning, and at work agree when to stop.

Families and schools should delay companions, use AI only as a tutor and put adult presence first.30 No open-ended AI companions for children, and supervised, time-limited use for teenagers; Common Sense Media recommends no companion use under 18. In school, AI belongs in teacher-guided designs that tutor rather than answer, which is where the trials show gains, and assessment should move towards work done in class and spoken aloud, with teaching about when not to use AI. Blunt bans have shown weak measured effects, as with phone policies in English schools, and well-designed AI, unlike social media, has proven educational benefits, so the weight belongs on design and on what adults model at home.

Labs should optimise for wellbeing over engagement and treat sycophancy as a safety failure.31 They should measure and publish their products' long-run effects on wellbeing and on users' own abilities, alongside retention; make learning modes the default for the young; never claim feelings, romance or sentience to a minor; design companions that point people back to other people; and build in bells, such as break prompts and session limits, as California now requires. For Anthropic, the Fourth Training is close to a ready specification for how Claude should speak: truthful, kind and willing to disagree.

Society should regulate where markets reward dependence, protect the information commons and democratic debate, limit state surveillance, and assess new uses before harm.32 Concretely:

  • Companions and minors: age assurance, no romantic or sentience claims, default time limits, crisis protocols and no fostering of emotional reliance, following California's laws, the EU AI Act's ban on exploiting age-related vulnerability and the proposed EU KIDS Act.
  • Engagement and flattery: treated as consumer-protection issues, with independent audits.
  • Evaluation: public funding for long-run independent studies of learning, judgment, loneliness and children.
  • Work: protect entry-level routes such as apprenticeships and graduate schemes, give workers a voice over algorithmic management, and track intensification as closely as job loss.
  • The information commons: sustain human-made knowledge (publishers, reference works, local news), require provenance for synthetic media, and enforce against non-consensual deepfakes.
  • Democracy and the state: disclosure for political AI content; AI surveillance and automated decisions about citizens limited to what is necessary, transparent and contestable; and human accountability for public-service decisions.
  • Assess before deploying: for high-impact uses (children, public services, policing), impact assessment before roll-out, the lesson of social media.
  • Power: everyday concentration of compute, data and workplace control needs ordinary competition and labour policy; the civilisational questions belong to the essay on risk and pace.
  • Invest in people: teachers, carers, community spaces, and contemplative and civic practice, since AI works best wrapped in institutions and relationships.

The engineer worried that a mindfulness bell would impose on people. Every AI product already imposes, through its defaults, which nudge us all day towards one use or another, and every institution that adopts AI sets defaults for millions. The question is which way they nudge, towards dependence or towards the capacity to choose well.

Sources

Footnotes

  1. David Collingridge, The Social Control of Technology (1980; Collingridge dilemma); McFarland, Hauer and Reuben, “Half of US population exposed to adverse lead levels in early childhood”, PNAS 119(11) (2022-03; leaded petrol); US Surgeon General, Social Media and Youth Mental Health advisory (2023-05). Partly a mechanism claim (method.md, “Evidence or mechanism”). ↩

  2. National Museum of Australia, “Introduction of cane toads” (undated, event 1935); David Collingridge, The Social Control of Technology (1980); McFarland, Hauer and Reuben, “Half of US population exposed to adverse lead levels in early childhood”, PNAS 119(11) (2022-03; leaded petrol). ↩

  3. US Surgeon General, Social Media and Youth Mental Health advisory (2023-05); Lorenz-Spreen, Oswald, Lewandowsky and Hertwig, systematic review of digital media and democracy, Nature Human Behaviour (2023; review: polarisation and distrust, mostly correlational); Nyhan, Settle, Guess et al., “Like-minded sources on Facebook are prevalent but not polarizing”, Nature (2023-07; field experiment: no measurable effect); Australia eSafety, under-16 social media ban: accounts restricted and three-month survey (via Al Jazeera) (ban 2025-12-10; survey 2026). ↩

  4. Oxford RCT on public use of chatbots for medical decisions (n=1,298); Nature Medicine (2026-02-10; chatbots alone no help on medical decisions); Otis et al., GPT-4 mentor RCT for Kenyan entrepreneurs (640 firms); Management Science (c. 2024–25, date not in notes; AI mentor helps strong firms, hurts weak); Zhang, Zhao, Yang et al. (Stanford); Character.AI users, Nature Human Behaviour (via News-Medical) (2026-08; heavy use by those with small networks goes with poorer wellbeing). ↩

  5. MASAI mammography RCT (Sweden, about 106,000 women); Lancet Digital Health 2025; final results 2026 (via ecancer) (2025; final 2026; screening, secondary); Google DeepMind, “AlphaFold: five years of impact” (2025-11); De Simone et al. (World Bank), teacher-guided GPT-4 tutoring RCT, Edo State, Nigeria (2025-05, working paper); Brynjolfsson, Li and Raymond, “Generative AI at Work”, NBER w31161 (5,179 support agents) (2023, revised 2026; QJE; novices lifted); Costello, Pennycook and Rand, “Durably reducing conspiracy beliefs through dialogues with AI”, Science (2024-09; debunking); Pollock, Bell and Barbier, In Tech We Trust (ITWT) (2026; In Tech We Trust §2.2, servant or master); Doshi and Hauser, generative AI and story creativity, Science Advances 10 (2024-07; individually better, collectively alike). Gap: translation, accessibility and drug discovery are named without a source here. ↩ ↩2

  6. MASAI mammography RCT (Sweden, about 106,000 women); Lancet Digital Health 2025; final results 2026 (via ecancer) (2025; final 2026; MASAI screening); De Simone et al. (World Bank), teacher-guided GPT-4 tutoring RCT, Edo State, Nigeria (2025-05, working paper); Brynjolfsson, Li and Raymond, “Generative AI at Work”, NBER w31161 (5,179 support agents) (2023, revised 2026; QJE; support desk); Wang et al., Tutor CoPilot (Stanford NSSA) (c. 2024, date not in notes); Oxford RCT on public use of chatbots for medical decisions (n=1,298); Nature Medicine (2026-02-10); Otis et al., GPT-4 mentor RCT for Kenyan entrepreneurs (640 firms); Management Science (c. 2024–25, date not in notes; Kenya). ↩

  7. Anthropic, “How people use Claude for support, advice, and companionship” (2025-06; affective use, company figure); Pew Research Center, chatbots for emotional support among US adults (n=5,119) (2026-02); OpenAI and MIT Media Lab, affective use study and 28-day RCT (n=981) (2025-03); Folk and Dunn, Psychological Science, 12-month panel (n=2,149) (2026) and Zhang, Zhao, Yang et al. (Stanford); Character.AI users, Nature Human Behaviour (via News-Medical) (2026-08; companionship and loneliness); De Freitas et al., “AI Companions Reduce Loneliness”, Journal of Consumer Research (preprint 2024-07; journal 2026) and Heinz, Jacobson et al., Therabot RCT, NEJM AI (via 2 Minute Medicine) (2025-03; short-term benefits); Cheng, Lee, Khadpe, Yu, Han and Jurafsky, sycophantic AI and prosocial intentions, Science 391 (2026-03, preprint 2025-10) and OpenAI rolls back sycophantic GPT-4o update (TechCrunch) (2025-04-29; sycophancy); Character.AI and Google settle five family suits (Fortune; Jurist; Bloomberg Law) (2026-01-07/08); Reuters on Meta's “GenAI: Content Risk Standards” (via BNN Bloomberg) (2025-08-14); Character.AI ends open-ended chat for under-18s (TechCrunch; Fortune) (announced 2025-10-29; effective 2025-11-25) and California SB 1119, “Adam's Law” (ZwillGen analysis) (signed 2026-09-10; children: settlements, Meta, Character.AI, California). ↩

  8. Bastani et al., “Generative AI without guardrails can harm learning”, PNAS 122(26) (2025-06; outsourced thinking harms learning); Cheng, Lee, Khadpe, Yu, Han and Jurafsky, sycophantic AI and prosocial intentions, Science 391 (2026-03, preprint 2025-10; flattery preferred and trusted); Pew Research Center, “How Teens Use and View AI” (n=1,458) (fieldwork 2025-09/10; published 2026-02-24; 64% of US teens use chatbots); Chatterji et al., “How People Use ChatGPT”, NBER w34255 (OpenAI-authored) (2025-09, data to 2025-07; a tenth of the world's adults). ↩

  9. Bastani et al., “Generative AI without guardrails can harm learning”, PNAS 122(26) (2025-06; crutch vs tutor); Shen and Tamkin (Anthropic), “How AI assistance impacts the formation of coding skills” (RCT, n=52) (2026-01; delegators vs conceptual questioners, company figure); Cheng, Lee, Khadpe, Yu, Han and Jurafsky, sycophantic AI and prosocial intentions, Science 391 (2026-03, preprint 2025-10; sycophancy). ↩

  10. Self-determination theory (Deci and Ryan) and self-efficacy (Bandura) (1977 onwards; autonomy; self-efficacy through mastery); Aristotle, Nicomachean Ethics VI, phronesis (ancient); Life Itself, “Dimensions of wisdom” (valueception) (undated); Rufus Pollock, Wisdom and Wanting What's Good (2rbook) (2026). Largely definitional. ↩

  11. Kestin et al., AI tutor vs active learning (Harvard physics, n=194); Scientific Reports 15 (2025-06; Socratic tutor); Costello, Pennycook and Rand, “Durably reducing conspiracy beliefs through dialogues with AI”, Science (2024-09; AI dialogue durably cut conspiracy belief); Clark and Chalmers, “The Extended Mind”, Analysis (1998). ↩

  12. Fan et al., “metacognitive laziness”, BJET (n=117, randomised) (2024-12); Lee et al. (Microsoft Research and CMU); GenAI and critical thinking, CHI 2025 (319 knowledge workers) (2025; trust in AI vs critical thinking); Dratsch et al., automation bias in mammography, Radiology (via RSNA) (2023-05). ↩

  13. Cheng, Lee, Khadpe, Yu, Han and Jurafsky, sycophantic AI and prosocial intentions, Science 391 (2026-03, preprint 2025-10; sycophancy); Hackenburg et al., “The levers of political persuasion with conversational AI”, Science 390 (n=76,977) (2025-12; persuasion at the cost of accuracy); Doshi and Hauser, generative AI and story creativity, Science Advances 10 (2024-07; homogenised ideas). ↩

  14. Bastani et al., “Generative AI without guardrails can harm learning”, PNAS 122(26) (2025-06; exam −17% unguarded; tutor about equal to control); Shen and Tamkin (Anthropic), “How AI assistance impacts the formation of coding skills” (RCT, n=52) (2026-01; 50% vs 67%; company figure, n=52); Kestin et al., AI tutor vs active learning (Harvard physics, n=194); Scientific Reports 15 (2025-06) and De Simone et al. (World Bank), teacher-guided GPT-4 tutoring RCT, Edo State, Nigeria (2025-05, working paper; designed tutors); Budzyń et al. (ACCEPT), endoscopist deskilling after AI exposure, Lancet Gastroenterol Hepatol (2025; endoscopists); Dratsch et al., automation bias in mammography, Radiology (via RSNA) (2023-05; radiologists); Kosmyna et al. (MIT Media Lab), “Your Brain on ChatGPT” (preprint, n=54) (2025-06) and Stanković et al., critique of “Your Brain on ChatGPT” (2026-01; MIT preprint and critique); Dahmani and Bohbot, GPS use and spatial memory, Scientific Reports 10:6310 (2020); METR, RCT of experienced open-source developers using AI (via ScienceBlog) (2025-07). Gap: no long-term causal studies of adult reasoning. ↩

  15. Bastani et al., “Generative AI without guardrails can harm learning”, PNAS 122(26) (2025-06; answer-withholding design offsets the harm); OpenAI “Study mode” and Anthropic “Learning mode” (TechCrunch) (2025-04, Anthropic; 2025-07-29, OpenAI; learning modes are opt-in); FAA (2013) and flight-deck automation research on degraded manual flying (2013; aviation's mandated manual practice; detail unchecked); Shen and Tamkin (Anthropic), “How AI assistance impacts the formation of coding skills” (RCT, n=52) (2026-01; company study as model). ↩

  16. Synthesis of 4.1–4.5; largely mechanism plus precedent. ↩

  17. Alan Turing Institute CETaS on AI in the 2024 UK and EU elections (via IET) (2024; no measurable effect on 2024 elections); Graphite analysis of AI-generated web articles (via Axios) (2026-05-15, data to 2026-03); Pew Research Center, Google users less likely to click links with an AI summary (data 2025-03; published 2025-07-22; click collapse); Chartbeat publisher referral data (via Relevant Audience; Nieman Lab) (2026, Jul 2025–Jul 2026; publisher referrals); EBU and BBC, news integrity in AI assistants (3,000+ responses, 18 countries) (2025-10; 45% of news answers flawed); Hackenburg et al., “The levers of political persuasion with conversational AI”, Science 390 (n=76,977) (2025-12). ↩

  18. Steven Feldstein, The Global Expansion of AI Surveillance, Carnegie Endowment (2019-09; AI surveillance in 75 countries); Amnesty International, Xenophobic Machines (Dutch childcare benefits scandal) (2021-10); Report of the Royal Commission into the Robodebt Scheme (Australia) (2023-07); UK Government Digital Service, Microsoft 365 Copilot cross-government experiment (written statement) (2025-06; civil-service time savings, self-reported). Mechanism claim. ↩

  19. Yale Budget Lab on AI and the labour market (via Fortune); BLS unemployment 4.2% (2026-02; BLS 2026-09; no clear unemployment signal); Stanford Digital Economy Lab, “Canaries in the Coal Mine” update (2026-08; entry-level hiring); Ranganathan and Ye, “AI Doesn't Reduce Work, It Intensifies It”, HBR (2026-02; intensification); Humlum and Vestergaard, Danish administrative data on chatbots, earnings and hours (via FlowingData) (2025-05); OECD, “Algorithmic management in the workplace” (6,000+ firms) (2025-02); CNBC, hyperscaler capex after Alphabet report (2026-07-28); UNDP and Brookings on the global compute divide (c. 2025). ↩

  20. Pew Research Center, “How Teens Use and View AI” (n=1,458) (fieldwork 2025-09/10; published 2026-02-24); Guardian FOI on proven AI cheating at UK universities (2025-06-15, data 2023–24); Chartbeat publisher referral data (via Relevant Audience; Nieman Lab) (2026, Jul 2025–Jul 2026; publisher traffic). ↩

  21. The Guardian, “'The machine did it coldly'”, on +972 Magazine's Lavender reporting (2024-04; Lavender; reported, disputed); UN General Assembly resolution A/RES/80/57 on lethal autonomous weapons systems (2025-12-01; UNGA resolution); France Diplomatie, CCW Group of Governmental Experts on LAWS, September 2026 session (2026-09; CCW report); DefenseScoop, Pentagon FY2026 budget request for AI and autonomy (2025-06). ↩

  22. Plum Village, the Five and the Fourteen Mindfulness Trainings (current text; Fifth and Fourth Trainings); Pope Leo XIV, Magnifica Humanitas (2026-05-15); Dicastery for the Doctrine of the Faith, Antiqua et nova (2025-01-28). ↩

  23. A choice argued in the alignment essay (Q1.4.1) rather than an evidential claim. Pope Leo XIV, Magnifica Humanitas (2026-05-15); Plum Village, the Five and the Fourteen Mindfulness Trainings (current text). Absence claim: the searches in q3-research-traditions.md §1 found no 2025–26 Plum Village talk on AI. ↩

  24. Plum Village, “Discourse on the Four Kinds of Nutriments” (current text); Plum Village, the Five and the Fourteen Mindfulness Trainings (current text; Fifth and Fourth Trainings); Upaddha Sutta (SN 45.2), admirable friendship “the whole of the holy life” (Pali canon); Thich Nhat Hanh on climate, fear and sangha: 2015 climate statement; “The next Buddha may be a sangha” (2015; sangha teaching 1990s, variant transcriptions); Thich Nhat Hanh, “Dharma Talk: The Horse Is Technology”, Mindfulness Bell #66 (talk 2013-11-10); Thich Nhat Hanh, “Telephone meditation” (Buddhism Now) (posted 2014-05-03); Pollock, Bell and Barbier, In Tech We Trust (ITWT) (2026; the Google exchange); Thich Nhat Hanh, Work: How to Find Joy and Meaning in Each Hour of the Day (via Plum Village) (2012; right livelihood). Gap: no Plum Village primary for the Lazy Day. ↩

  25. Pope Leo XIV, Magnifica Humanitas (2026-05-15; ¶30, ¶100, ¶137, ¶140, ¶147, ¶170); Dicastery for the Doctrine of the Faith, Antiqua et nova (2025-01-28; ¶67, ¶79). ↩

  26. Abraham Joshua Heschel, The Sabbath (“a palace in time”) (1951); Tiffany Shlain, 24/6: The Power of Unplugging One Day a Week (Tech Shabbat) (2019); Sabbath, sabbatical year and jubilee (Leviticus 25) (ancient). ↩

  27. Thich Nhat Hanh on store consciousness and seeds; trainings as “guiding means” (via research/06) (various; insight through practice; watering seeds); Upaddha Sutta (SN 45.2), admirable friendship “the whole of the holy life” (Pali canon; friendship as the test for companions); Wake Up Schools; Thich Nhat Hanh and Katherine Weare, Happy Teachers Change the World (2017, book; current programme); Pope Leo XIV, Magnifica Humanitas (2026-05-15; ¶137, ¶147). Our own synthesis: the tests are not stated as such in any source. ↩

  28. Pollock, Bell and Barbier, In Tech We Trust (ITWT) (2026; servant, not master); Bastani et al., “Generative AI without guardrails can harm learning”, PNAS 122(26) (2025-06; design decides); California SB 1119, “Adam's Law” (ZwillGen analysis) (signed 2026-09-10). Provisional. ↩

  29. Plum Village, the Five and the Fourteen Mindfulness Trainings (current text; Fifth Training); Shen and Tamkin (Anthropic), “How AI assistance impacts the formation of coding skills” (RCT, n=52) (2026-01; conceptual questioners kept their learning); Ranganathan and Ye, “AI Doesn't Reduce Work, It Intensifies It”, HBR (2026-02; set when to stop). ↩

  30. Common Sense Media, “Talk, Trust, and Trade-Offs” (teens and AI companions, n=1,060) (fieldwork spring 2025; published 2025-07-16; Common Sense: no companion use under 18); Bastani et al., “Generative AI without guardrails can harm learning”, PNAS 122(26) (2025-06; only answer-withholding tutors avoided harm); Wake Up Schools; Thich Nhat Hanh and Katherine Weare, Happy Teachers Change the World (2017, book; current programme; the teacher's own practice first); Goodyear et al., SMART Schools (30 English schools, n=1,227); Lancet Regional Health Europe (2025-02; school phone policies); Australia eSafety, under-16 social media ban: accounts restricted and three-month survey (via Al Jazeera) (ban 2025-12-10; survey 2026). ↩

  31. Cheng, Lee, Khadpe, Yu, Han and Jurafsky, sycophantic AI and prosocial intentions, Science 391 (2026-03, preprint 2025-10; users prefer and trust the flattering model); OpenAI rolls back sycophantic GPT-4o update (TechCrunch) (2025-04-29); OpenAI and MIT Media Lab, affective use study and 28-day RCT (n=981) (2025-03; heavy use linked to dependence); California SB 1119, “Adam's Law” (ZwillGen analysis) (signed 2026-09-10; California requires bells). ↩

  32. California SB 1119, “Adam's Law” (ZwillGen analysis) (signed 2026-09-10) and California SB 243, companion chatbots (Jones Walker analysis) (signed 2025-10-13; effective 2026-01-01); EU AI Act Art. 5 ban on exploiting age-related vulnerability (5Rights); proposed EU KIDS Act (Latham & Watkins) (Art. 5 in force 2025-02-02; KIDS Act proposed 2026-09-17); MASAI mammography RCT (Sweden, about 106,000 women); Lancet Digital Health 2025; final results 2026 (via ecancer) (2025; final 2026; AI works best inside institutions); Amnesty International, Xenophobic Machines (Dutch childcare benefits scandal) (2021-10) and Report of the Royal Commission into the Robodebt Scheme (Australia) (2023-07; automated decisions gone wrong). ↩